nb_kernel104_ia32_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,968 行 · 第 1/5 页
S
1,968 行
movlhps %xmm2,%xmm0 ## xmm0= jxH1a jxH1b jxH1c jxH1d movaps %xmm0,nb104_jxH1(%esp) movhlps %xmm6,%xmm2 ## xmm2= jyH1a jyH1b jyH1c jyH1d movaps %xmm2,nb104_jyH1(%esp) movlhps %xmm3,%xmm1 movaps %xmm1,nb104_jxH2(%esp) movhlps %xmm7,%xmm3 movaps %xmm4,%xmm6 movaps %xmm3,nb104_jyH2(%esp) movlhps %xmm5,%xmm4 movaps %xmm4,nb104_jxM(%esp) movhlps %xmm6,%xmm5 movaps %xmm5,nb104_jyM(%esp) movss 20(%esi,%eax,4),%xmm0 movss 32(%esi,%eax,4),%xmm1 movss 44(%esi,%eax,4),%xmm2 movss 20(%esi,%ecx,4),%xmm3 movss 32(%esi,%ecx,4),%xmm4 movss 44(%esi,%ecx,4),%xmm5 movhps 16(%esi,%ebx,4),%xmm0 movhps 28(%esi,%ebx,4),%xmm1 movhps 40(%esi,%ebx,4),%xmm2 movhps 16(%esi,%edx,4),%xmm3 movhps 28(%esi,%edx,4),%xmm4 movhps 40(%esi,%edx,4),%xmm5 shufps $204,%xmm3,%xmm0 ## constant 11001100 shufps $204,%xmm4,%xmm1 ## constant 11001100 shufps $204,%xmm5,%xmm2 ## constant 11001100 movaps %xmm0,nb104_jzH1(%esp) movaps %xmm1,nb104_jzH2(%esp) movaps %xmm2,nb104_jzM(%esp) movaps nb104_ixH1(%esp),%xmm0 movaps nb104_iyH1(%esp),%xmm1 movaps nb104_izH1(%esp),%xmm2 movaps nb104_ixH1(%esp),%xmm3 movaps nb104_iyH1(%esp),%xmm4 movaps nb104_izH1(%esp),%xmm5 subps nb104_jxH1(%esp),%xmm0 subps nb104_jyH1(%esp),%xmm1 subps nb104_jzH1(%esp),%xmm2 subps nb104_jxH2(%esp),%xmm3 subps nb104_jyH2(%esp),%xmm4 subps nb104_jzH2(%esp),%xmm5 movaps %xmm0,nb104_dxH1H1(%esp) movaps %xmm1,nb104_dyH1H1(%esp) movaps %xmm2,nb104_dzH1H1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb104_dxH1H2(%esp) movaps %xmm4,nb104_dyH1H2(%esp) movaps %xmm5,nb104_dzH1H2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb104_rsqH1H1(%esp) movaps %xmm3,nb104_rsqH1H2(%esp) movaps nb104_ixH1(%esp),%xmm0 movaps nb104_iyH1(%esp),%xmm1 movaps nb104_izH1(%esp),%xmm2 movaps nb104_ixH2(%esp),%xmm3 movaps nb104_iyH2(%esp),%xmm4 movaps nb104_izH2(%esp),%xmm5 subps nb104_jxM(%esp),%xmm0 subps nb104_jyM(%esp),%xmm1 subps nb104_jzM(%esp),%xmm2 subps nb104_jxH1(%esp),%xmm3 subps nb104_jyH1(%esp),%xmm4 subps nb104_jzH1(%esp),%xmm5 movaps %xmm0,nb104_dxH1M(%esp) movaps %xmm1,nb104_dyH1M(%esp) movaps %xmm2,nb104_dzH1M(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb104_dxH2H1(%esp) movaps %xmm4,nb104_dyH2H1(%esp) movaps %xmm5,nb104_dzH2H1(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb104_rsqH1M(%esp) movaps %xmm3,nb104_rsqH2H1(%esp) movaps nb104_ixH2(%esp),%xmm0 movaps nb104_iyH2(%esp),%xmm1 movaps nb104_izH2(%esp),%xmm2 movaps nb104_ixH2(%esp),%xmm3 movaps nb104_iyH2(%esp),%xmm4 movaps nb104_izH2(%esp),%xmm5 subps nb104_jxH2(%esp),%xmm0 subps nb104_jyH2(%esp),%xmm1 subps nb104_jzH2(%esp),%xmm2 subps nb104_jxM(%esp),%xmm3 subps nb104_jyM(%esp),%xmm4 subps nb104_jzM(%esp),%xmm5 movaps %xmm0,nb104_dxH2H2(%esp) movaps %xmm1,nb104_dyH2H2(%esp) movaps %xmm2,nb104_dzH2H2(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb104_dxH2M(%esp) movaps %xmm4,nb104_dyH2M(%esp) movaps %xmm5,nb104_dzH2M(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm4,%xmm3 addps %xmm5,%xmm3 movaps %xmm0,nb104_rsqH2H2(%esp) movaps %xmm3,nb104_rsqH2M(%esp) movaps nb104_ixM(%esp),%xmm0 movaps nb104_iyM(%esp),%xmm1 movaps nb104_izM(%esp),%xmm2 movaps nb104_ixM(%esp),%xmm3 movaps nb104_iyM(%esp),%xmm4 movaps nb104_izM(%esp),%xmm5 subps nb104_jxH1(%esp),%xmm0 subps nb104_jyH1(%esp),%xmm1 subps nb104_jzH1(%esp),%xmm2 subps nb104_jxH2(%esp),%xmm3 subps nb104_jyH2(%esp),%xmm4 subps nb104_jzH2(%esp),%xmm5 movaps %xmm0,nb104_dxMH1(%esp) movaps %xmm1,nb104_dyMH1(%esp) movaps %xmm2,nb104_dzMH1(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 movaps %xmm3,nb104_dxMH2(%esp) movaps %xmm4,nb104_dyMH2(%esp) movaps %xmm5,nb104_dzMH2(%esp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm1,%xmm0 addps %xmm2,%xmm0 addps %xmm3,%xmm4 addps %xmm5,%xmm4 movaps %xmm0,nb104_rsqMH1(%esp) movaps %xmm4,nb104_rsqMH2(%esp) movaps nb104_ixM(%esp),%xmm0 movaps nb104_iyM(%esp),%xmm1 movaps nb104_izM(%esp),%xmm2 subps nb104_jxM(%esp),%xmm0 subps nb104_jyM(%esp),%xmm1 subps nb104_jzM(%esp),%xmm2 movaps %xmm0,nb104_dxMM(%esp) movaps %xmm1,nb104_dyMM(%esp) movaps %xmm2,nb104_dzMM(%esp) mulps %xmm0,%xmm0 mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 addps %xmm1,%xmm0 addps %xmm2,%xmm0 movaps %xmm0,nb104_rsqMM(%esp) ## start doing invsqrt use rsq values in xmm0, xmm4 rsqrtps %xmm0,%xmm1 rsqrtps %xmm4,%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb104_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps %xmm0,%xmm1 mulps %xmm4,%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb104_half(%esp),%xmm3 ## rinvMM mulps nb104_half(%esp),%xmm7 ## rinvMH2 movaps %xmm3,nb104_rinvMM(%esp) movaps %xmm7,nb104_rinvMH2(%esp) rsqrtps nb104_rsqH1H1(%esp),%xmm1 rsqrtps nb104_rsqH1H2(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb104_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb104_rsqH1H1(%esp),%xmm1 mulps nb104_rsqH1H2(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb104_half(%esp),%xmm3 mulps nb104_half(%esp),%xmm7 movaps %xmm3,nb104_rinvH1H1(%esp) movaps %xmm7,nb104_rinvH1H2(%esp) rsqrtps nb104_rsqH1M(%esp),%xmm1 rsqrtps nb104_rsqH2H1(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb104_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb104_rsqH1M(%esp),%xmm1 mulps nb104_rsqH2H1(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb104_half(%esp),%xmm3 mulps nb104_half(%esp),%xmm7 movaps %xmm3,nb104_rinvH1M(%esp) movaps %xmm7,nb104_rinvH2H1(%esp) rsqrtps nb104_rsqH2H2(%esp),%xmm1 rsqrtps nb104_rsqH2M(%esp),%xmm5 movaps %xmm1,%xmm2 movaps %xmm5,%xmm6 mulps %xmm1,%xmm1 mulps %xmm5,%xmm5 movaps nb104_three(%esp),%xmm3 movaps %xmm3,%xmm7 mulps nb104_rsqH2H2(%esp),%xmm1 mulps nb104_rsqH2M(%esp),%xmm5 subps %xmm1,%xmm3 subps %xmm5,%xmm7 mulps %xmm2,%xmm3 mulps %xmm6,%xmm7 mulps nb104_half(%esp),%xmm3 mulps nb104_half(%esp),%xmm7 movaps %xmm3,nb104_rinvH2H2(%esp) movaps %xmm7,nb104_rinvH2M(%esp) rsqrtps nb104_rsqMH1(%esp),%xmm1 movaps %xmm1,%xmm2 mulps %xmm1,%xmm1 movaps nb104_three(%esp),%xmm3 mulps nb104_rsqMH1(%esp),%xmm1 subps %xmm1,%xmm3 mulps %xmm2,%xmm3 mulps nb104_half(%esp),%xmm3 movaps %xmm3,nb104_rinvMH1(%esp) ## start with H1-H1 interaction movaps nb104_rinvH1H1(%esp),%xmm0 movaps %xmm0,%xmm7 mulps %xmm0,%xmm0 mulps nb104_qqHH(%esp),%xmm7 mulps %xmm7,%xmm0 addps nb104_vctot(%esp),%xmm7 movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb104_dxH1H1(%esp),%xmm0 mulps nb104_dyH1H1(%esp),%xmm1 mulps nb104_dzH1H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixH1(%esp),%xmm0 addps nb104_fiyH1(%esp),%xmm1 addps nb104_fizH1(%esp),%xmm2 movaps %xmm3,nb104_fjxH1(%esp) movaps %xmm4,nb104_fjyH1(%esp) movaps %xmm5,nb104_fjzH1(%esp) movaps %xmm0,nb104_fixH1(%esp) movaps %xmm1,nb104_fiyH1(%esp) movaps %xmm2,nb104_fizH1(%esp) ## H1-H2 interaction movaps nb104_rinvH1H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H1-H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb104_dxH1H2(%esp),%xmm0 mulps nb104_dyH1H2(%esp),%xmm1 mulps nb104_dzH1H2(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixH1(%esp),%xmm0 addps nb104_fiyH1(%esp),%xmm1 addps nb104_fizH1(%esp),%xmm2 movaps %xmm3,nb104_fjxH2(%esp) movaps %xmm4,nb104_fjyH2(%esp) movaps %xmm5,nb104_fjzH2(%esp) movaps %xmm0,nb104_fixH1(%esp) movaps %xmm1,nb104_fiyH1(%esp) movaps %xmm2,nb104_fizH1(%esp) ## H1-M interaction movaps nb104_rinvH1M(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqMH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H1-M addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 xorps %xmm3,%xmm3 movaps %xmm3,%xmm4 movaps %xmm3,%xmm5 mulps nb104_dxH1M(%esp),%xmm0 mulps nb104_dyH1M(%esp),%xmm1 mulps nb104_dzH1M(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixH1(%esp),%xmm0 addps nb104_fiyH1(%esp),%xmm1 addps nb104_fizH1(%esp),%xmm2 movaps %xmm3,nb104_fjxM(%esp) movaps %xmm4,nb104_fjyM(%esp) movaps %xmm5,nb104_fjzM(%esp) movaps %xmm0,nb104_fixH1(%esp) movaps %xmm1,nb104_fiyH1(%esp) movaps %xmm2,nb104_fizH1(%esp) ## H2-H1 interaction movaps nb104_rinvH2H1(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fs H2-H1 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb104_fjxH1(%esp),%xmm3 movaps nb104_fjyH1(%esp),%xmm4 movaps nb104_fjzH1(%esp),%xmm5 mulps nb104_dxH2H1(%esp),%xmm0 mulps nb104_dyH2H1(%esp),%xmm1 mulps nb104_dzH2H1(%esp),%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 addps nb104_fixH2(%esp),%xmm0 addps nb104_fiyH2(%esp),%xmm1 addps nb104_fizH2(%esp),%xmm2 movaps %xmm3,nb104_fjxH1(%esp) movaps %xmm4,nb104_fjyH1(%esp) movaps %xmm5,nb104_fjzH1(%esp) movaps %xmm0,nb104_fixH2(%esp) movaps %xmm1,nb104_fiyH2(%esp) movaps %xmm2,nb104_fizH2(%esp) ## H2-H2 interaction movaps nb104_rinvH2H2(%esp),%xmm0 movaps %xmm0,%xmm1 mulps %xmm0,%xmm0 mulps nb104_qqHH(%esp),%xmm1 mulps %xmm1,%xmm0 ## fsH2H2 addps %xmm1,%xmm7 ## add to local vctot movaps %xmm0,%xmm1 movaps %xmm0,%xmm2 movaps nb104_fjxH2(%esp),%xmm3 movaps nb104_fjyH2(%esp),%xmm4 movaps nb104_fjzH2(%esp),%xmm5 mulps nb104_dxH2H2(%esp),%xmm0 mulps nb104_dyH2H2(%esp),%xmm1 mulps nb104_dzH2H2(%esp),%xmm2
⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?